An open-source TypeScript framework for agentic apps: define an action once and it is exposed to the UI, the agent, an HTTP API, a CLI, MCP and A2A, with auth, permissions and real-time sync built in. Verdict: a strong pick for React/TypeScript teams who want an agent and a real UI sharing the same actions and state; it is beta, so expect API churn.
Checks turns CI into a pre-commit check that developers and coding agents can call from the chainctl CLI, a git hook, the GitHub App or MCP, running in warm, credential-free microVMs and returning a signed verdict. Workspaces gives each developer or agent session a hardened microVM dev environment with short-lived injected credentials. Verdict: aimed at teams running autonomous coding agents who want a real CI gate before an agent opens a PR; both are beta with waitlists.
inclusionAI published the weights of Ling-3.1-flash on Hugging Face under the MIT license: a hybrid-reasoning MoE with 560B total and 25B active parameters and a 1M-token context, with SGLang and vLLM recipes. Verdict: one of the largest permissively licensed long-context models you can self-host, but the reference setups need 8x H200-class GPUs, so most developers will use it through an API provider.
OpenTelemetry tracing on the Claude Platform (public beta)
Claude Managed Agents, the Messages API and the Claude SDKs can now export standard OpenTelemetry traces (agent turns, model calls, tool and MCP calls, with latency, tokens and IDs) to any OTLP tool such as Grafana, Honeycomb, Sentry, Datadog, Langfuse or LangSmith. Verdict: if you ship Claude-based agents, this replaces hand-rolled instrumentation and puts agent steps in the same trace as the rest of your stack.
Google's new frontier model is now available to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search, and to developers on the paid Gemini API.
The open-source agent's biggest release adds a guided first-run setup on a free trial model, hosted app connectors alongside MCP servers, one-click local models via a llama.cpp runtime, a bot screen you can take over, and SemVer stable releases. Verdict: the easiest way yet to try a fully open, model-agnostic personal agent, and the local-model path is a real plus for privacy-minded developers.